Probabilistic forecasting of water level in the Beijing-Hangzhou Grand Canal based on MC dropout and CNN-LSTM

ABSTRACT

Reliable water level forecasting with quantified uncertainty is vital for urban flood control operations. This study proposes a hybrid probabilistic water level forecasting framework integrating CNN, LSTM, and Monte Carlo Dropout to address the risk-sensitive operation of urban flood control infrastructure. The model utilizes a heteroscedastic Gaussian likelihood loss to simultaneously capture point predictions and quantify predictive uncertainty. Validated with hourly data (2020–2024) from the Beijing–Hangzhou Grand Canal, the framework achieves an RMSE of 0.016 m, R2 of 0.995, and a PICP of 0.951 at a 1-h lead time, matching deterministic benchmarks while providing well-calibrated uncertainty estimates. Feature importance analysis identifies distinct predictors across hydrological regimes. The framework maintains robust performance up to 24 h (R2 = 0.959, PICP = 0.957), delivering high-precision forecasts and quantified uncertainty essential for modern urban water management.